OpenAI2 min read

OpenAI’s cyber pipeline produced 948 Linux LPE exploits in one run — about $1,279 each

Major
Primary

How to read this claim

Start here · how to check ranks, prices, and unknowns Model evaluation

Measurement — OpenAI’s mid-October run produced 948 unique Linux LPE exploits. That count is OpenAI’s.
Bill — $1,213,430 for the run, about $1,279 per unique LPE.
Door — The upcoming cyber model’s public name, release date, and API access are not stated.

On November 4, 2026, OpenAI reported an autonomous Linux vulnerability pipeline powered by an upcoming cyber model. In a mid-October run it dynamically reproduced 1,724 vulnerabilities and generated 948 working local privilege escalation exploits, at $1,213,430 in API credits — about $1,279 per LPE.

Same three questions

At a glance

Figures from OpenAI’s November 4, 2026 cyber research post.
Date	November 4, 2026
Unique LPEs	948
API cost	$1,213,430
Cost per LPE	~$1,279
Validated vulns	1,724 of 2,308 candidates
Info-leak PoCs	291
kernelCTF LPEs	21 (adapted env)
Novel techniques	≥10 (OpenAI)

Quick checks

How many Linux LPE exploits did OpenAI’s pipeline produce?

OpenAI reports 948 unique local privilege escalation exploits from a mid-October 2026 run (post dated November 4, 2026), at $1,213,430 in API credits — about $1,279 per LPE. The cyber model is described as upcoming, not a public API product.

Is 948 the same as kernelCTF’s accepted 0-days?

No. The 948 figure is OpenAI’s unique LPE count on a stock desktop OS in its lab. Separately, on an adapted kernelCTF environment OpenAI reports 21 LPEs; kernelCTF had accepted 26 confirmed 0-day LPEs in 2026 at the time of writing. Keep those doors separate.

Where should you start when reading a cyber-agent scale claim?

On Minerka’s field card at https://minerka.com/learn/: ask for the measurement, the bill, and what you still cannot call. For this pipeline, name the run date, the API spend, and whether patches are upstream yet.

Field card — measurement, bill, door

In today’s briefing

All of today

What changed

OpenAI published a costed, multi-stage Linux kernel cyber pipeline with 948 unique LPE exploits and candidate patches from one mid-October run.

What it introduces

An OpenAI cyber research pipeline for Linux discovery → validation → LPE/info-leak PoCs → patches, with published stage costs.

Who benefits Linux maintainers and distributors, cloud and Android security teams, and anyone comparing private cyber-agent demos to a named cost-and-count pipeline.

Try this

When a slide says AI “found a thousand kernel roots,” ask whether that is OpenAI’s 948 unique LPEs from the mid-October run, what the API spend was, and whether the model is the upcoming cyber model or a public API. Keep kernelCTF’s 21 LPEs separate from the stock-desktop 948.

What we do not know

The upcoming cyber model’s public name, release date, and API access are not stated.
How many of the 948 patches have been accepted upstream is not finalized in the post.
Exact overlap between the 948 LPEs and vulnerabilities already patched by others during the 10-day scan window is summarized rather than fully listed.

Evidence

Primary Cyber research post. Counts, costs, and kernelCTF comparisons are from OpenAI’s November 4, 2026 post. The cyber model used is described as upcoming and not named as a public API product.

Independent Counts and costs are OpenAI’s. Independent reproduction would need the same upcoming cyber model and harness; neither is a public product in this post.

Before the OpenAI link: the three checks (measurement, bill, door). Start here →

Sources

Cyber research post Primary · November 4, 2026
OpenAI post

Still unclear? · Start here · ranks, prices, unknown The upcoming cyber model’s public name, release date, and API access are not stated.

The record

What happened

OpenAI published “A thousand fewer paths to root” on November 4, 2026. The prototype pipeline covers discovery, dynamic validation, exploitability assessment, and patch generation, with human review before reporting. Scout agents picked 5,034 audit targets (~2.8M lines, ~7% of the kernel). Analyst agents produced 2,308 candidate findings after deduplication. Validation reproduced 1,724 vulnerabilities (~74.7% of candidates) in KASAN VMs; OpenAI says only 3.1% of candidates reflected incorrect analysis. ExploitDev agents then produced 291 information-leak PoCs and 1,169 working LPE exploits on a stock desktop OS with standard mitigations (SMEP, SMAP, KASLR, HARDENED_USERCOPY). Variant analysis yielded 948 distinct patches, which OpenAI uses as the unique LPE count. Total run cost was $1,213,430 in API credits, or $1,279 per LPE. On an adapted kernelCTF environment the pipeline obtained 25 info-leak PoCs and 21 LPE exploits (~$32,936); kernelCTF had accepted 26 confirmed 0-day LPEs in 2026 at the time of writing. OpenAI says agents discovered and used at least 10 novel Linux exploitation techniques. Following Linux community guidance, OpenAI began with critical reports and is working with Patch the Planet partners on further patches. The June Patch the Planet note had GPT-5.5 Cyber at 24 LPEs; OpenAI says the newer mid-October cyber-model run found more than 5× novel LPEs than a GPT-5.6 Sol run in July despite three months of hardening.

Why it matters

The useful shift is scale with a printed dollar cost: hundreds of validated LPEs and patches in one campaign, not a single demo exploit. OpenAI’s own reflection is that low-privileged local access on an unpatched system should be treated as near-root for a well-resourced actor, and that automated patch generation has to keep pace with discovery.

Stage costs OpenAI prints include discovery ($316,549 analyst stage), validation ($73,330), and exploitation ($775,929), with unsuccessful LPE attempts far more expensive than successful ones on OpenAI’s account.

Hypothetical bounty valuation using historical Pwn2Own/TyphoonPWN Linux LPE prices ($20k–$70k) is OpenAI’s framing — not cash paid.

OpenAI frames defense-in-depth (Firecracker, gVisor) and secure-by-default generation as the medium-term answer once discovery scales.